MeSH Embeddings

This repository contains code & pre-trained representations for the Medical Subject Headings (MeSH) thesaurus. These representations were trained using the node2vec algorithm with default parameters. For more details about node2vec please visit this repository.

Note: node2vec relies on an edge list to learn node representations. The edge list for MeSH can be constructed using the tree numbers from the xml file (descYYYY.xml) which is available here. In order to enable anybody to train their own MeSH representations, the edge list for desc2020.xml is shared along with the pre-trained representations.

The code below shows how to use the vectors in practice. Two files are needed

• the pre-trained vectors: mesh_embeddings.txt.gz
• the mapping between MeSH Unique Identifiers and vector ids: mesh_ui_to_id.pickle

import gzip
import pickle

with open('mesh_ui_to_id.pickle', 'rb') as stream:
    mesh_ui_to_id = pickle.load(stream)

embeddings = {}
with gzip.open('mesh_embeddings.txt.gz', 'rt') as stream:
    n_embeddings, embedding_dim = stream.readline().strip().split()
    for line in stream:
        splitline = str(line).strip().split()
        idx = int(splitline[0])
        vector = list(map(float, splitline[1:]))
        embeddings[idx] = vector

def get_embedding_from_mesh_ui(mesh_ui):
    return embeddings[mesh_ui_to_id[mesh_ui]]

print(f'There are {n_embeddings} MeSH embeddings')
print(f'Each embedding is {embedding_dim}-dimensional')

print('MeSH UI for <Headache> is: D006261')
print('MeSH embedding for <Headache> is:', get_embedding_from_mesh_ui('D006261'))

>>> There are 29638 MeSH embeddings
>>> Each embedding is 256-dimensional
>>> MeSH UI for <Headache> is: D006261
>>> MeSH embedding for <Headache> is: [-0.22856602, -0.32223737, -0.16364807, ...]